Re: Connection Header

Jeffrey Mogul writes:
>My own intuition is that we can follow two paths:
>	(1) Stick with the parallel connection approach, which gives
>	reasonably good performance today but which might turn into
>	a global performance disaster in the future.
>	(2) Encourage people to shift ASAP to an alternative (besides
>	the WIDTH and HEIGHT tags, several HTTP-level mechanisms have
>	been proposed).  This might avert future global problems, but
>	in the short term it might be a while before enough servers
>	support it to make it highly beneficial.  I would also expect
>	that in the long run (i.e., with widespread support), it would
>	improve UPP because you wouldn't have to wait even for the initial
>	bits of the images to arrive.

If we go with something like an HTTP-based image hinting scheme, I don't
think the rate of adoption is a big an issue as you imply.  A browser like
Netscape could continue to have multiple parallel connections to older
servers if it wanted to.  Path (1) as you've described is really not a path
at all - it's just staying with the limits of what we have today.

I think the most important point you make is that an image hinting approach
will have a greatly higher UPP than a parallel connection scheme, especially
over very slow connections.  The way I see it, there are two main goals:

1) Have the formatted text appear as quickly as possible.
2) Have the entire document with pictures appear as quickly as possible.

Every image that's being downloaded takes away bandwidth from the text.  For
the first goal, image hinting on a single connection clearly wins over
parallel connections.

The parallel connection scheme helps somewhat with the second goal, since
the RTTs are occuring simultaneously.  However, keeping the connection alive
and using an MGET for the retrieval would remove most of the RTT penalty
anyway.  Again, even if the adoption of these enhancements were slow, a
client would still have the option of multiple connections.

Also, my belief is that defining the extensions and getting them into
browsers will propel server authors to add them fairly quickly.  The high
resource utilization of an http server being subjected to parallel
connection clients have been discussed in some depth on c.i.w.providers.
Many providers would jump on a single-connection approach just for server
efficiency reasons.

--
Jim Seidman
Senior Software Engineer
Spyglass, Inc.

Received on Thursday, 22 December 1994 07:24:47 UTC